[erzaehlsalex]

What

Adds ModelCapabilities.input with maxImages and an optional maxMedia. Each chat model now reports the most images it accepts in one request, so apps can trim or batch images before sending instead of finding out from ImageCountExceededError (#229).

Like output.maxTokens, the field is present only when the provider has a verified limit for the model id on the host it serves. If it is missing, the limit is unknown. It does not mean there is no limit.

Provider	maxImages	Source
Anthropic, Anthropic on Vertex	600 for 1M-context models, 100 for 200k models; images and PDF pages share the limit, so maxMedia is set too	platform.claude.com vision and context-windows docs; matches the captured … > 100 errors
OpenAI (Responses and Chat Completions)	1,500 for image-capable models	developers.openai.com images-and-vision guide
Gemini API	3,600 for known Gemini models	ai.google.dev image-understanding docs; matches the captured error
Vertex AI Gemini	3,000 for known Gemini models	every Vertex AI model page
Mistral	8 for Mistral's vision models	docs.mistral.ai vision FAQ; matches the captured rejection
Azure OpenAI, OpenAI-compatible, Omnifact, xAI, Kimi	not set	Azure's docs contradict each other (50 vs 10). The other hosts enforce their own limits; xAI and Kimi document none.

Design notes

• Expose only, no client-side check. Providers also count document pages and video frames against the limit, so counting locally would disagree with them. ImageCountExceededError handling stays as it is.
• Unknown ids get no limit. This covers fail-open registries, self-hosted endpoints and image-generation models, which accept far fewer images (Gemini Flash Image allows 14).
• Google platform. getGoogleModelCapabilities(modelId, platform) and a platform factory option on createGoogleGenAIProvider. @core-ai/google-vertex passes 'vertex', and its getGoogleModelCapabilities export now returns Vertex AI's limits.
• OpenAI wrappers. When a provider has no registry of its own, the shared factory falls back to OpenAI's registry, and that fallback now applies omitOpenAIInputLimits. This covers Azure v1, openai-compat and Omnifact. Azure classic applies it explicitly.

Not covered

• Mistral's 8 is not live-probed for each family. One Mistral docs page says the limit "depends on the model".
• Only image counts are covered. Per-image byte and pixel limits are out of scope.

Testing

• npm run build, turbo lint check-types test, npm run docs:check-links
• New capability tests for each provider, plus provider tests checking that Vertex reports 3,000 and that Azure (v1 and classic), openai-compat and Omnifact report no limit.

🤖 Generated with Claude Code

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit ebdf83a.

Approve preview

[erzaehlsalex]

We can add 600 and 100 into constants